That would create a cottage industry of borderline-infringing works designed to trigger the filter but win in court.

I.e. they would trigger bad copyright takedowns that weren't legally justified, forcing companies to stop doing those.

If it's possible to trigger those intentionally, then it's likely that it's also happening accidentally, and this policy would make that stop.

Possibly, but machine generating edge cases for the purpose of collecting fines feels like it should also be wrong.

True, but those edge cases have to convince a court made up of humans, who will probably have similar feelings. That might be a decent check on abuses.

I’m no lawyer, but deliberately performing adversarial attacks to get rich from being automatically and incorrectly fined feels like it ought to be somewhere between fraud and hacking (in the Computer Misuse Act sense rather than thr hackernews sense).